Researchers propose LineageRAG for smarter AI evidence retrieval
New GraphRAG method improves multi-hop question answering by 5+ points
A team of researchers from universities including Peking University and the University of Science and Technology of China has proposed LineageRAG, a novel approach to GraphRAG that explicitly constructs evidence lineages with source grounding.
The method initializes query-derived evidence demands and expands each lineage through demand-conditioned retrieval over corpus graphs, retaining provenance for every candidate. When evidence supports a demand, the system grounds it in verbatim source text, creating transparent connections between questions and answers. Experiments on HotpotQA, 2WikiMultiHopQA, and MuSiQue show average improvements of 3.51 points in recall@5, 5.96 points in exact match accuracy, and 5.22 points in F1 score over leading GraphRAG baselines. This addresses a key limitation in existing GraphRAG systems where evidence discovery and source grounding were implicitly connected.
- LineageRAG improves GraphRAG performance by explicitly constructing evidence lineages with source grounding
- Achieves average gains of 3.51 (R@5), 5.96 (EM), and 5.22 (F1) points over baselines on HotpotQA and similar datasets
- Method creates transparent provenance chains between queries and evidence sources
Why It Matters
Enables more accurate enterprise RAG systems with verifiable multi-hop question answering capabilities